OpenKB memory profile

Type: agentic-system-analyses/types/agentic-system-memory-profile.md

Comparison rationale

This profile classifies accepted records, not target execution. All implemented findings retain a wired basis; the Workbench human-edit and external-host navigation alternatives retain afforded basis. No supplied execution or causal evidence upgrades those findings. The inventory includes local exported artifacts and unresolved PageIndex persistence, but does not turn provider model weights into accumulated OpenKB memory.

The sole reconciliation supersession replaces the epistemic source-content object with MEM-OBJ-source-content. All source-artifact findings therefore cite that canonical declaration. Summary, concept and entity records remain valid parts of RT-OBJ-wiki-pages; they are not superseded. Source prose and images, metadata, catalog entries, saved answers, session display and reusable SDK history remain distinct operative parts.

Compilation supports consolidation and revision without establishing semantic novelty. Full-body replacement is an implemented evolve operation; requests for cross-document synthesis do not establish a synthesize finding without recorded claim-level evidence. Source removal changes availability and associations but does not establish history-preserving invalidation, temporal decay or complete semantic forgetting. Hash duplicate skips and graph edge uniqueness are not memory dedup transformations. See EPI-RTE-source-withdrawal, MEM-OBJ-hashes and MEM-OBJ-graph.

Authority follows the consumer rather than the file extension. Retained source/body text supplies knowledge; compact descriptions and recurrence cues influence planner priority; source lists, pointers and registry entries route work. Hash entries also enforce duplicate rejection. Saved evaluation labels supply validation references and select coverage questions, not semantic acceptance of an installed skill. Instruction candidates in output/skills are not automatically installed guidance. These distinctions follow MEM-OBJ-page-frontmatter, MEM-OBJ-eval-sets and RT-BAP-skill-instructions.

Requested file/tool retrieval is pull. Automatic composition of retained history, compact planner context and local instruction catalogs is push. Session IDs and update slugs select actual retained parts; the mere presence of a path on a requested file read does not establish identifier push. No save-to-index guarantee is inferred for explorations. See RT-RTE-session-chat, RT-RTE-document-add and MEM-OBJ-explorations.

The trace-learning positives concern durable automatic context writes, not demonstrated improved capacity: interaction history is sanitized and later prepended, and conversational answers/notes can be retained for later reads. History retains tool records as well as conversational turns; the browser display trace is not its substitute. Document-to-wiki-to-skill generation alone establishes neither experience learning nor its absence. No evidenced-absence record supports a local no on the remaining inventory. Thus the learning union has supported yes with partial coverage, and trace-source coverage remains partial rather than inapplicable. See MEM-OBJ-session-history, MEM-OBJ-session-display and MEM-OBJ-explorations.

Missing initial instruction provenance/control, private PageIndex state semantics, actual script contents, semantic novelty and unrecorded alternative trace inputs prevent stronger coverage assessments. They do not erase supported local findings or convert implementation into observed benefit. No additional source reads were made; all warrants are in the supplied accepted records and their annotations.